Fairness Enforcement in Unlabeled Data Classifiers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning systems face challenges in ensuring fairness and accuracy in decision-making processes, particularly in domains like credit and employment, where bias in training data can lead to biased results.
Innovation Solution
The approach involves generating a trained classifier with enforced fairness by creating a data set with known characteristics, including label bias, feature rarity discrepancy, and selection bias, and then using this data set to train the classifier alongside additional unlabeled data points and specific training parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If fairness constraints are enforced during training or postprocessing, then fairness is improved, but accuracy decreases
Solution Approach 1:
The patent applies preliminary action by generating synthetic training data with known fairness characteristics before actual model training. This synthetic data is created with controlled label bias, selection bias, and feature rarity parameters to pre-establish a foundation for fair learning, allowing the model to learn fairness constraints before encountering real biased data.
Solution Approach 2:
The patent utilizes parameter changes by systematically varying fairness parameters (label bias α, selection bias β, feature rarity γ) during synthetic data generation and training. By adjusting these parameters and observing model performance, the system identifies optimal parameter settings that achieve both fairness and accuracy, transforming the fixed fairness-accuracy tradeoff into a tunable optimization problem.
2Measurement precision
If training data is used as-is from real-world sources, then accuracy may be maintained, but bias in results increases
Solution Approach 1:
The patent applies copying by creating synthetic copies of real-world training data with controlled fairness properties. Instead of directly using potentially biased real-world data, the system generates synthetic data that replicates the statistical characteristics and complexity of real data while allowing precise control over fairness parameters, enabling fair training without sacrificing realistic data characteristics.
Solution Approach 2:
The patent introduces synthetic data as an intermediary between real-world biased data and the training process. This intermediary layer allows the model to learn from data with known fairness characteristics before being deployed on real data, acting as a buffer that mediates the transfer of knowledge while filtering out unwanted biases.
3Reliability
If existing fairness enforcement methods are applied, then some fairness is achieved, but they fail to account for selection bias and feature rarity discrepancies
Solution Approach 1:
The patent achieves universality by creating a comprehensive fairness framework that simultaneously handles multiple types of bias (label bias, selection bias, feature rarity bias) within a single unified approach. The synthetic data generation process incorporates all three bias types with controllable parameters, and the training methodology addresses them collectively, making the solution applicable to a wide range of biased datasets regardless of which specific bias is present.
Data Source
AI summary
Fairness of a trained classifier may be ensured by generating a data set for training, the data set generated using input data points of a feature space including multiple dimensions and according to different parameters including an amount of label bias, a control for discrepancy between rarity of features, and an amount of selection bias. Unlabeled data points of the input data comprising unobserved ground truths are labeled according to the amount of label bias and the input data sampled according to the amount of selection bias and the control for the discrepancy between the rarity of features. The classifier is then trained using the sampled and labeled data points as well as additional 10 unlabeled data points. The trained classifier is then usable to determine unbiased classifications of one or more labels for one or more other data sets.